Improving E-commerce search queries using transformer-based named entity recognition
Knowledge-Based Systems, cilt.346, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 346
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.knosys.2026.116213
- Dergi Adı: Knowledge-Based Systems
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Library, Information Science & Technology Abstracts (LISTA)
- Anahtar Kelimeler: E-commerce search queries, Named entity recognition (NER), Natural language processing (NLP), Transformers-based models
- Marmara Üniversitesi Adresli: Evet
Özet
Named Entity Recognition (NER) is crucial for enhancing product discovery in e-commerce. In this study, we introduce a novel, fine-grained annotated dataset for Turkish e-commerce search queries and systematically evaluate baseline and transformer-based models. Quantitative results show that transformers significantly outperform traditional baselines, with ELECTRA and a hybrid BERT+Bi-LSTM+CRF model achieving peak weighted F1-scores of 91.97% and 92.49%, respectively. Beyond architectural comparisons, we conduct an empirical morphological evaluation and a zero-shot Large Language Model (LLM) benchmark. Our morphological analysis reveals that agglutinative suffixation in Turkish increases semantic density, providing stronger contextual cues that actively improve NER performance. Furthermore, we benchmarked BERT — selected for its optimal balance of accuracy and computational efficiency — against a zero-shot LLM (Gemini Pro). While fine-tuned transformers yield higher F1-scores, qualitative analysis indicates they often overfit to ground truth noise and annotator bias. Conversely, the LLM exhibits superior zero-shot robustness to typographical errors and semantic ambiguities. Ultimately, this study presents highly efficient NER models while offering deep linguistic insights for LLM-based research in low-resource agglutinative languages.